ADR-1026: R6 SYCL kernel correctness — rd-stride OOB and unchecked graph_wait¶

Context¶

Round-6 analysis found two classes of correctness bugs in the SYCL feature extractor paths:

VIF rd-downsample stride truncation for odd widths (two locations): integer_vif_sycl.cpp lines 822 and 1313 use e_w / 2 as the rd-buffer row stride. When e_w is odd (e.g. 1279 pixels), integer division truncates: 1279 / 2 = 639, but the downsampled row has (1279 + 1) / 2 = 640 pixels. The last active thread on that row writes to rd_ref[gy/2 * 639 + 639] which is one element into the next row, corrupting the adjacent row's data. The allocation uses the rounded-up formula (w+1)/2 * (h+1)/2; the stride must match.
Unchecked vmaf_sycl_graph_wait() return values (three locations): integer_motion_sycl.cpp:610, integer_vif_sycl.cpp:1635, and integer_adm_sycl.cpp:1559 all call vmaf_sycl_graph_wait() (which returns int) and discard the result. If the SYCL queue throws (device fault or reset), the wait returns a non-zero error code; the extractor then reads stale accumulator memory and emits a wrong score — silently, with no error propagation to the caller.

Replace e_w / 2 with (e_w + 1u) / 2u for the rd stride in both VIF SYCL variants (standard SIMD and SIMD-16).
Capture the vmaf_sycl_graph_wait() return value in all three collect functions and propagate it as an early return.

Option	Pros	Cons	Why not chosen
Assert `e_w % 2 == 0` at launch time	Catches odd widths earlier	Rejects valid odd-width content	VIF must handle arbitrary resolutions

Positive: SYCL VIF scores are correct for odd-width inputs; device-fault errors on the SYCL path are now propagated to the scoring pipeline instead of silently producing stale scores.
Negative: A device fault now surfaces as a scoring error rather than a stale-but-numeric result. Callers that did not check the collect return value before will now see non-zero returns.
Neutral: No change to the score values for even-width inputs.